/// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> static public IList <int> Candidates( IList <int> candidate_items, CandidateItems candidate_item_mode, IPosOnlyFeedback test, IPosOnlyFeedback training) { switch (candidate_item_mode) { case CandidateItems.TRAINING: return(training.AllItems); case CandidateItems.TEST: return(test.AllItems); case CandidateItems.OVERLAP: var result = test.AllItems.Intersect(training.AllItems).ToList(); result.Shuffle(); return(result); case CandidateItems.UNION: result = test.AllItems.Union(training.AllItems).ToList(); result.Shuffle(); return(result); case CandidateItems.EXPLICIT: if (candidate_items == null) { throw new ArgumentNullException("candidate_items"); } return(candidate_items); default: throw new ArgumentException("Unknown candidate_item_mode: " + candidate_item_mode.ToString()); } }
protected override void LoadData() { base.LoadData(); // test users if (test_users_file != null) test_users = user_mapping.ToInternalID( File.ReadLines(Path.Combine(data_dir, test_users_file)).ToArray() ); else test_users = test_data != null ? test_data.AllUsers : training_data.AllUsers; // candidate items if (candidate_items_file != null) candidate_items = item_mapping.ToInternalID( File.ReadLines(Path.Combine(data_dir, candidate_items_file)).ToArray() ); else if (all_items) candidate_items = Enumerable.Range(0, item_mapping.InternalIDs.Max() + 1).ToArray(); if (candidate_items != null) eval_item_mode = CandidateItems.EXPLICIT; else if (in_training_items) eval_item_mode = CandidateItems.TRAINING; else if (in_test_items) eval_item_mode = CandidateItems.TEST; else if (overlap_items) eval_item_mode = CandidateItems.OVERLAP; else eval_item_mode = CandidateItems.UNION; }
/// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> public static IList<int> Candidates( IList<int> candidate_items, CandidateItems candidate_item_mode, IPosOnlyFeedback test, IPosOnlyFeedback training) { switch (candidate_item_mode) { case CandidateItems.TRAINING: return training.AllItems; case CandidateItems.TEST: return test.AllItems; case CandidateItems.OVERLAP: var result = test.AllItems.Intersect(training.AllItems).ToList(); result.Shuffle(); return result; case CandidateItems.UNION: result = test.AllItems.Union(training.AllItems).ToList(); result.Shuffle(); return result; case CandidateItems.EXPLICIT: if (candidate_items == null) throw new ArgumentNullException("candidate_items"); return candidate_items; default: throw new ArgumentException("Unknown candidate_item_mode: " + candidate_item_mode.ToString()); } }
/// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> static public IList<int> Candidates( IList<int> candidate_items, CandidateItems candidate_item_mode, IPosOnlyFeedback test, IPosOnlyFeedback training) { IList<int> test_items = (test != null) ? test.AllItems : new int[0]; IList<int> result = null; switch (candidate_item_mode) { case CandidateItems.TRAINING: result = training.AllItems.ToArray(); break; case CandidateItems.TEST: result = test.AllItems.ToArray(); break; case CandidateItems.OVERLAP: result = test_items.Intersect(training.AllItems).ToList(); break; case CandidateItems.UNION: result = test_items.Union(training.AllItems).ToList(); break; case CandidateItems.EXPLICIT: if (candidate_items == null) throw new ArgumentNullException("candidate_items"); result = candidate_items.ToArray(); break; default: throw new ArgumentException("Unknown candidate_item_mode: " + candidate_item_mode.ToString()); } result.Shuffle(); return result; }
/// <summary>Computes the AUC fit of a recommender on the training data</summary> /// <returns>the AUC on the training data</returns> /// <param name='recommender'>the item recommender to evaluate</param> /// <param name="test_users">a list of integers with all test users; if null, use all users in the test cases</param> /// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> public static double ComputeFit( this ItemRecommender recommender, IList<int> test_users = null, IList<int> candidate_items = null, CandidateItems candidate_item_mode = CandidateItems.OVERLAP) { return recommender.Evaluate(recommender.Feedback, recommender.Feedback, test_users, candidate_items, candidate_item_mode, true)["RMSE"]; }
/// <summary>Online evaluation for rankings of items</summary> /// <remarks> /// The evaluation protocol works as follows: /// For every test user, evaluate on the test items, and then add the those test items to the training set and perform an incremental update. /// The sequence of users is random. /// </remarks> /// <param name="recommender">the item recommender to be evaluated</param> /// <param name="test">test cases</param> /// <param name="training">training data (must be connected to the recommender's training data)</param> /// <param name="test_users">a list of all test user IDs</param> /// <param name="candidate_items">a list of all candidate item IDs</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <returns>a dictionary containing the evaluation results (averaged by user)</returns> public static ItemRecommendationEvaluationResults EvaluateOnline( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode) { var incremental_recommender = recommender as IIncrementalItemRecommender; if (incremental_recommender == null) throw new ArgumentException("recommender must be of type IIncrementalItemRecommender"); // prepare candidate items once to avoid recreating them switch (candidate_item_mode) { case CandidateItems.TRAINING: candidate_items = training.AllItems; break; case CandidateItems.TEST: candidate_items = test.AllItems; break; case CandidateItems.OVERLAP: candidate_items = new List<int>(test.AllItems.Intersect(training.AllItems)); break; case CandidateItems.UNION: candidate_items = new List<int>(test.AllItems.Union(training.AllItems)); break; } test_users.Shuffle(); var results_by_user = new Dictionary<int, ItemRecommendationEvaluationResults>(); foreach (int user_id in test_users) { if (candidate_items.Intersect(test.ByUser[user_id]).Count() == 0) continue; // prepare data var current_test_data = new PosOnlyFeedback<SparseBooleanMatrix>(); foreach (int index in test.ByUser[user_id]) current_test_data.Add(user_id, test.Items[index]); // evaluate user var current_result = Items.Evaluate(recommender, current_test_data, training, current_test_data.AllUsers, candidate_items, CandidateItems.EXPLICIT); results_by_user[user_id] = current_result; // update recommender var tuples = new List<Tuple<int, int>>(); foreach (int index in test.ByUser[user_id]) tuples.Add(Tuple.Create(user_id, test.Items[index])); incremental_recommender.AddFeedback(tuples); } var results = new ItemRecommendationEvaluationResults(); foreach (int u in results_by_user.Keys) foreach (string measure in Items.Measures) results[measure] += results_by_user[u][measure]; foreach (string measure in Items.Measures) results[measure] /= results_by_user.Count; results["num_users"] = results_by_user.Count; results["num_items"] = candidate_items.Count; results["num_lists"] = results_by_user.Count; return results; }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public EvaluationResults DoRatingBasedRankingCrossValidation( this RatingPredictor recommender, uint num_folds, IList<int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var split = new RatingCrossValidationSplit(recommender.Ratings, num_folds); return recommender.DoRatingBasedRankingCrossValidation(split, candidate_items, candidate_item_mode, compute_fit, show_results); }
/// <summary>Computes the AUC fit of a recommender on the training data</summary> /// <returns>the AUC on the training data</returns> /// <param name='recommender'>the item recommender to evaluate</param> /// <param name="test_users">a list of integers with all test users; if null, use all users in the test cases</param> /// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> public static double ComputeFit( this ItemRecommender recommender, IList <int> test_users = null, IList <int> candidate_items = null, CandidateItems candidate_item_mode = CandidateItems.OVERLAP) { return(recommender.Evaluate( recommender.Feedback, recommender.Feedback, test_users, candidate_items, candidate_item_mode, RepeatedEvents.Yes)["AUC"]); }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public EvaluationResults DoRatingBasedRankingCrossValidation( this RatingPredictor recommender, uint num_folds, IList <int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var split = new RatingCrossValidationSplit(recommender.Ratings, num_folds); return(recommender.DoRatingBasedRankingCrossValidation(split, candidate_items, candidate_item_mode, compute_fit, show_results)); }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a dataset split</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> public static ItemRecommendationEvaluationResults DoCrossValidation( this IRecommender recommender, ISplit<IPosOnlyFeedback> split, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var avg_results = new ItemRecommendationEvaluationResults(); if (!(recommender is ItemRecommender)) throw new ArgumentException("recommender must be of type ItemRecommender"); Parallel.For(0, (int) split.NumberOfFolds, fold => { try { var split_recommender = (ItemRecommender) recommender.Clone(); // avoid changes in recommender split_recommender.Feedback = split.Train[fold]; split_recommender.Train(); var fold_results = Items.Evaluate(split_recommender, split.Test[fold], split.Train[fold], test_users, candidate_items, candidate_item_mode); if (compute_fit) fold_results["fit"] = (float) split_recommender.ComputeFit(); // thread-safe stats lock (avg_results) foreach (var key in fold_results.Keys) if (avg_results.ContainsKey(key)) avg_results[key] += fold_results[key]; else avg_results[key] = fold_results[key]; if (show_results) Console.Error.WriteLine("fold {0} {1}", fold, fold_results); } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (var key in Items.Measures) avg_results[key] /= split.NumberOfFolds; avg_results["num_users"] /= split.NumberOfFolds; avg_results["num_items"] /= split.NumberOfFolds; if (compute_fit) avg_results["fit"] /= split.NumberOfFolds; return avg_results; }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a dataset split</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public EvaluationResults DoRatingBasedRankingCrossValidation( this RatingPredictor recommender, ISplit<IRatings> split, IList<int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var avg_results = new ItemRecommendationEvaluationResults(); Parallel.For(0, (int) split.NumberOfFolds, fold => { try { var split_recommender = (RatingPredictor) recommender.Clone(); // avoid changes in recommender split_recommender.Ratings = split.Train[fold]; split_recommender.Train(); var test_data_posonly = new PosOnlyFeedback<SparseBooleanMatrix>(split.Test[fold]); var training_data_posonly = new PosOnlyFeedback<SparseBooleanMatrix>(split.Train[fold]); IList<int> test_users = test_data_posonly.AllUsers; var fold_results = Items.Evaluate(split_recommender, test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode); if (compute_fit) fold_results["fit"] = (float) split_recommender.ComputeFit(); // thread-safe stats lock (avg_results) foreach (var key in fold_results.Keys) if (avg_results.ContainsKey(key)) avg_results[key] += fold_results[key]; else avg_results[key] = fold_results[key]; if (show_results) Console.Error.WriteLine("fold {0} {1}", fold, fold_results); } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (var key in Items.Measures) avg_results[key] /= split.NumberOfFolds; avg_results["num_users"] /= split.NumberOfFolds; avg_results["num_items"] /= split.NumberOfFolds; return avg_results; }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> static public void DoRatingBasedRankingIterativeCrossValidation( this RatingPredictor recommender, uint num_folds, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { var split = new RatingCrossValidationSplit(recommender.Ratings, num_folds); recommender.DoRatingBasedRankingIterativeCrossValidation(split, test_users, candidate_items, candidate_item_mode, repeated_events, max_iter, find_iter); }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> public static ItemRecommendationEvaluationResults DoCrossValidation( this IRecommender recommender, uint num_folds, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { if (!(recommender is ItemRecommender)) throw new ArgumentException("recommender must be of type ItemRecommender"); var split = new PosOnlyFeedbackCrossValidationSplit<PosOnlyFeedback<SparseBooleanMatrix>>(((ItemRecommender) recommender).Feedback, num_folds); return recommender.DoCrossValidation(split, test_users, candidate_items, candidate_item_mode, compute_fit, show_results); }
protected override void LoadData() { base.LoadData(); // test users if (test_users_file != null) { test_users = user_mapping.ToInternalID(File.ReadLines(Path.Combine(data_dir, test_users_file)).ToArray()); } else { test_users = test_data != null ? test_data.AllUsers : training_data.AllUsers; } // candidate items if (candidate_items_file != null) { candidate_items = item_mapping.ToInternalID(File.ReadLines(Path.Combine(data_dir, candidate_items_file)).ToArray()); } else if (all_items) { candidate_items = Enumerable.Range(0, item_mapping.InternalIDs.Max() + 1).ToArray(); } if (candidate_items != null) { eval_item_mode = CandidateItems.EXPLICIT; } else if (in_training_items) { eval_item_mode = CandidateItems.TRAINING; } else if (in_test_items) { eval_item_mode = CandidateItems.TEST; } else if (overlap_items) { eval_item_mode = CandidateItems.OVERLAP; } else { eval_item_mode = CandidateItems.UNION; } }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public ItemRecommendationEvaluationResults DoCrossValidation( this IRecommender recommender, uint num_folds, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { if (!(recommender is ItemRecommender)) { throw new ArgumentException("recommender must be of type ItemRecommender"); } var split = new PosOnlyFeedbackCrossValidationSplit <PosOnlyFeedback <SparseBooleanMatrix> >(((ItemRecommender)recommender).Feedback, num_folds); return(recommender.DoCrossValidation(split, test_users, candidate_items, candidate_item_mode, compute_fit, show_results)); }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> static public void DoIterativeCrossValidation( this IRecommender recommender, uint num_folds, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { if (!(recommender is ItemRecommender)) { throw new ArgumentException("recommender must be of type ItemRecommender"); } var split = new PosOnlyFeedbackCrossValidationSplit <PosOnlyFeedback <SparseBooleanMatrix> >(((ItemRecommender)recommender).Feedback, num_folds); recommender.DoIterativeCrossValidation(split, test_users, candidate_items, candidate_item_mode, repeated_events, max_iter, find_iter); }
/// <summary>Evaluation for rankings of items</summary> /// <remarks> /// User-item combinations that appear in both sets are ignored for the test set, and thus in the evaluation, /// except the boolean argument repeated_events is set. /// /// The evaluation measures are listed in the Measures property. /// Additionally, 'num_users' and 'num_items' report the number of users that were used to compute the results /// and the number of items that were taken into account. /// /// Literature: /// <list type="bullet"> /// <item><description> /// C. Manning, P. Raghavan, H. Schütze: Introduction to Information Retrieval, Cambridge University Press, 2008 /// </description></item> /// </list> /// /// On multi-core/multi-processor systems, the routine tries to use as many cores as possible, /// which should to an almost linear speed-up. /// </remarks> /// <param name="recommender">item recommender</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> /// <param name="test_users">a list of integers with all test users; if null, use all users in the test cases</param> /// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="n">length of the item list to evaluate -- if set to -1 (default), use the complete list, otherwise compute evaluation measures on the top n items</param> /// <returns>a dictionary containing the evaluation results (default is false)</returns> static public ItemRecommendationEvaluationResults Evaluate( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList <int> test_users = null, IList <int> candidate_items = null, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, RepeatedEvents repeated_events = RepeatedEvents.No, int n = -1) { if (test_users == null) { test_users = test.AllUsers; } candidate_items = Candidates(candidate_items, candidate_item_mode, test, training); var result = new ItemRecommendationEvaluationResults(); // make sure that the user matrix is completely initialized before entering parallel code var training_user_matrix = training.UserMatrix; var test_user_matrix = test.UserMatrix; int num_users = 0; Parallel.ForEach(test_users, user_id => { try { var correct_items = new HashSet <int>(test_user_matrix[user_id]); correct_items.IntersectWith(candidate_items); if (correct_items.Count == 0) { return; } var ignore_items_for_this_user = new HashSet <int>( repeated_events == RepeatedEvents.Yes || training_user_matrix[user_id] == null ? new int[0] : training_user_matrix[user_id] ); ignore_items_for_this_user.IntersectWith(candidate_items); int num_candidates_for_this_user = candidate_items.Count - ignore_items_for_this_user.Count; if (correct_items.Count == num_candidates_for_this_user) { return; } var prediction = recommender.Recommend(user_id, candidate_items: candidate_items, n: n, ignore_items: ignore_items_for_this_user); var prediction_list = (from t in prediction select t.Item1).ToArray(); int num_dropped_items = num_candidates_for_this_user - prediction.Count; double auc = AUC.Compute(prediction_list, correct_items, num_dropped_items); double map = PrecisionAndRecall.AP(prediction_list, correct_items); double ndcg = NDCG.Compute(prediction_list, correct_items); double rr = ReciprocalRank.Compute(prediction_list, correct_items); var positions = new int[] { 5, 10 }; var prec = PrecisionAndRecall.PrecisionAt(prediction_list, correct_items, positions); var recall = PrecisionAndRecall.RecallAt(prediction_list, correct_items, positions); // thread-safe incrementing lock (result) { num_users++; result["AUC"] += (float)auc; result["MAP"] += (float)map; result["NDCG"] += (float)ndcg; result["MRR"] += (float)rr; result["prec@5"] += (float)prec[5]; result["prec@10"] += (float)prec[10]; result["recall@5"] += (float)recall[5]; result["recall@10"] += (float)recall[10]; } if (num_users % 1000 == 0) { Console.Error.Write("."); } if (num_users % 60000 == 0) { Console.Error.WriteLine(); } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (string measure in Measures) { result[measure] /= num_users; } result["num_users"] = num_users; result["num_lists"] = num_users; result["num_items"] = candidate_items.Count; return(result); }
/// <summary>Online evaluation for rankings of items</summary> /// <remarks> /// The evaluation protocol works as follows: /// For every test user, evaluate on the test items, and then add the those test items to the training set and perform an incremental update. /// The sequence of users is random. /// </remarks> /// <param name="recommender">the item recommender to be evaluated</param> /// <param name="test">test cases</param> /// <param name="training">training data (must be connected to the recommender's training data)</param> /// <param name="test_users">a list of all test user IDs</param> /// <param name="candidate_items">a list of all candidate item IDs</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <returns>a dictionary containing the evaluation results (averaged by user)</returns> static public ItemRecommendationEvaluationResults EvaluateOnline( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode) { var incremental_recommender = recommender as IIncrementalItemRecommender; if (incremental_recommender == null) { throw new ArgumentException("recommender must be of type IIncrementalItemRecommender"); } candidate_items = Items.Candidates(candidate_items, candidate_item_mode, test, training); test_users.Shuffle(); var results_by_user = new Dictionary <int, ItemRecommendationEvaluationResults>(); foreach (int user_id in test_users) { if (candidate_items.Intersect(test.ByUser[user_id]).Count() == 0) { continue; } // prepare data var current_test_data = new PosOnlyFeedback <SparseBooleanMatrix>(); foreach (int index in test.ByUser[user_id]) { current_test_data.Add(user_id, test.Items[index]); } // evaluate user var current_result = Items.Evaluate(recommender, current_test_data, training, current_test_data.AllUsers, candidate_items, CandidateItems.EXPLICIT); results_by_user[user_id] = current_result; // update recommender var tuples = new List <Tuple <int, int> >(); foreach (int index in test.ByUser[user_id]) { tuples.Add(Tuple.Create(user_id, test.Items[index])); } incremental_recommender.AddFeedback(tuples); // TODO candidate_items should be updated properly } var results = new ItemRecommendationEvaluationResults(); foreach (int u in results_by_user.Keys) { foreach (string measure in Items.Measures) { results[measure] += results_by_user[u][measure]; } } foreach (string measure in Items.Measures) { results[measure] /= results_by_user.Count; } results["num_users"] = results_by_user.Count; results["num_items"] = candidate_items.Count; results["num_lists"] = results_by_user.Count; return(results); }
/// <summary>Evaluation for rankings of items</summary> /// <remarks> /// User-item combinations that appear in both sets are ignored for the test set, and thus in the evaluation, /// except the boolean argument repeated_events is set. /// /// The evaluation measures are listed in the Measures property. /// Additionally, 'num_users' and 'num_items' report the number of users that were used to compute the results /// and the number of items that were taken into account. /// /// Literature: /// <list type="bullet"> /// <item><description> /// C. Manning, P. Raghavan, H. Schütze: Introduction to Information Retrieval, Cambridge University Press, 2008 /// </description></item> /// </list> /// /// On multi-core/multi-processor systems, the routine tries to use as many cores as possible, /// which should to an almost linear speed-up. /// </remarks> /// <param name="recommender">item recommender</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> /// <param name="test_users">a list of integers with all test users; if null, use all users in the test cases</param> /// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="n">length of the item list to evaluate -- if set to -1 (default), use the complete list, otherwise compute evaluation measures on the top n items</param> /// <returns>a dictionary containing the evaluation results (default is false)</returns> public static ItemRecommendationEvaluationResults Evaluate( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList<int> test_users = null, IList<int> candidate_items = null, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, RepeatedEvents repeated_events = RepeatedEvents.No, int n = -1) { if (test_users == null) test_users = test.AllUsers; candidate_items = Candidates(candidate_items, candidate_item_mode, test, training); var result = new ItemRecommendationEvaluationResults(); // make sure that the user matrix is completely initialized before entering parallel code var training_user_matrix = training.UserMatrix; var test_user_matrix = test.UserMatrix; int num_users = 0; Parallel.ForEach(test_users, user_id => { try { var correct_items = new HashSet<int>(test_user_matrix[user_id]); correct_items.IntersectWith(candidate_items); if (correct_items.Count == 0) return; var ignore_items_for_this_user = new HashSet<int>( repeated_events == RepeatedEvents.Yes || training_user_matrix[user_id] == null ? new int[0] : training_user_matrix[user_id] ); ignore_items_for_this_user.IntersectWith(candidate_items); int num_candidates_for_this_user = candidate_items.Count - ignore_items_for_this_user.Count; if (correct_items.Count == num_candidates_for_this_user) return; var prediction = recommender.Recommend(user_id, candidate_items:candidate_items, n:n, ignore_items:ignore_items_for_this_user); var prediction_list = (from t in prediction select t.Item1).ToArray(); int num_dropped_items = num_candidates_for_this_user - prediction.Count; double auc = AUC.Compute(prediction_list, correct_items, num_dropped_items); double map = PrecisionAndRecall.AP(prediction_list, correct_items); double ndcg = NDCG.Compute(prediction_list, correct_items); double rr = ReciprocalRank.Compute(prediction_list, correct_items); var positions = new int[] { 5, 10 }; var prec = PrecisionAndRecall.PrecisionAt(prediction_list, correct_items, positions); var recall = PrecisionAndRecall.RecallAt(prediction_list, correct_items, positions); // thread-safe incrementing lock (result) { num_users++; result["AUC"] += (float) auc; result["MAP"] += (float) map; result["NDCG"] += (float) ndcg; result["MRR"] += (float) rr; result["prec@5"] += (float) prec[5]; result["prec@10"] += (float) prec[10]; result["recall@5"] += (float) recall[5]; result["recall@10"] += (float) recall[10]; } if (num_users % 1000 == 0) Console.Error.Write("."); if (num_users % 60000 == 0) Console.Error.WriteLine(); } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (string measure in Measures) result[measure] /= num_users; result["num_users"] = num_users; result["num_lists"] = num_users; result["num_items"] = candidate_items.Count; return result; }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a dataset split</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public ItemRecommendationEvaluationResults DoCrossValidation( this IRecommender recommender, ISplit <IPosOnlyFeedback> split, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var avg_results = new ItemRecommendationEvaluationResults(); if (!(recommender is ItemRecommender)) { throw new ArgumentException("recommender must be of type ItemRecommender"); } Parallel.For(0, (int)split.NumberOfFolds, fold => { try { var split_recommender = (ItemRecommender)recommender.Clone(); // avoid changes in recommender split_recommender.Feedback = split.Train[fold]; split_recommender.Train(); var fold_results = Items.Evaluate(split_recommender, split.Test[fold], split.Train[fold], test_users, candidate_items, candidate_item_mode); if (compute_fit) { fold_results["fit"] = (float)split_recommender.ComputeFit(); } // thread-safe stats lock (avg_results) foreach (var key in fold_results.Keys) { if (avg_results.ContainsKey(key)) { avg_results[key] += fold_results[key]; } else { avg_results[key] = fold_results[key]; } } if (show_results) { Console.Error.WriteLine("fold {0} {1}", fold, fold_results); } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (var key in Items.Measures) { avg_results[key] /= split.NumberOfFolds; } avg_results["num_users"] /= split.NumberOfFolds; avg_results["num_items"] /= split.NumberOfFolds; return(avg_results); }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> static public void DoRatingBasedRankingIterativeCrossValidation( this RatingPredictor recommender, uint num_folds, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { var split = new RatingCrossValidationSplit(recommender.Ratings, num_folds); recommender.DoRatingBasedRankingIterativeCrossValidation(split, test_users, candidate_items, candidate_item_mode, repeated_events, max_iter, find_iter); }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a positive-only feedback dataset split</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> public static void DoIterativeCrossValidation( this IRecommender recommender, ISplit<IPosOnlyFeedback> split, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { if (!(recommender is IIterativeModel)) throw new ArgumentException("recommender must be of type IIterativeModel"); if (!(recommender is ItemRecommender)) throw new ArgumentException("recommender must be of type ItemRecommender"); var split_recommenders = new ItemRecommender[split.NumberOfFolds]; var iterative_recommenders = new IIterativeModel[split.NumberOfFolds]; var fold_results = new ItemRecommendationEvaluationResults[split.NumberOfFolds]; // initial training and evaluation Parallel.For(0, (int) split.NumberOfFolds, i => { try { split_recommenders[i] = (ItemRecommender) recommender.Clone(); // to avoid changes in recommender split_recommenders[i].Feedback = split.Train[i]; split_recommenders[i].Train(); iterative_recommenders[i] = (IIterativeModel) split_recommenders[i]; fold_results[i] = Items.Evaluate(split_recommenders[i], split.Test[i], split.Train[i], test_users, candidate_items, candidate_item_mode, repeated_events); if (show_fold_results) Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, iterative_recommenders[i].NumIter); } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), iterative_recommenders[0].NumIter); // iterative training and evaluation for (int it = (int) iterative_recommenders[0].NumIter + 1; it <= max_iter; it++) { Parallel.For(0, (int) split.NumberOfFolds, i => { try { iterative_recommenders[i].Iterate(); if (it % find_iter == 0) { fold_results[i] = Items.Evaluate(split_recommenders[i], split.Test[i], split.Train[i], test_users, candidate_items, candidate_item_mode, repeated_events); if (show_fold_results) Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, it); } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), it); } }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="num_folds">the number of folds</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> public static void DoIterativeCrossValidation( this IRecommender recommender, uint num_folds, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { if (!(recommender is ItemRecommender)) throw new ArgumentException("recommender must be of type ItemRecommender"); var split = new PosOnlyFeedbackCrossValidationSplit<PosOnlyFeedback<SparseBooleanMatrix>>(((ItemRecommender) recommender).Feedback, num_folds); recommender.DoIterativeCrossValidation(split, test_users, candidate_items, candidate_item_mode, repeated_events, max_iter, find_iter); }
protected override void LoadData() { TimeSpan loading_time = Wrap.MeasureTime(delegate() { base.LoadData(); // training data training_data = double.IsNaN(rating_threshold) ? ItemData.Read(training_file, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE) : ItemDataRatingThreshold.Read(training_file, rating_threshold, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE); // test data if (test_ratio == 0) { if (test_file != null) { test_data = double.IsNaN(rating_threshold) ? ItemData.Read(test_file, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE) : ItemDataRatingThreshold.Read(test_file, rating_threshold, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE); } } else { var split = new PosOnlyFeedbackSimpleSplit <PosOnlyFeedback <SparseBooleanMatrix> >(training_data, test_ratio); training_data = split.Train[0]; test_data = split.Test[0]; } if (user_prediction) { // swap file names for test users and candidate items var ruf = test_users_file; var rif = candidate_items_file; test_users_file = rif; candidate_items_file = ruf; // swap user and item mappings var um = user_mapping; var im = item_mapping; user_mapping = im; item_mapping = um; // transpose training and test data training_data = training_data.Transpose(); // transpose test data if (test_data != null) { test_data = test_data.Transpose(); } } for (int i = 0; i < recommenders.Count; i++) { if (recommenders[i] is MyMediaLite.ItemRecommendation.ItemRecommender) { ((ItemRecommender)recommenders[i]).Feedback = training_data; } } // test users if (test_users_file != null) { test_users = user_mapping.ToInternalID(File.ReadLines(Path.Combine(data_dir, test_users_file)).ToArray()); } else { test_users = test_data != null ? test_data.AllUsers : training_data.AllUsers; } // if necessary, perform user sampling if (num_test_users > 0 && num_test_users < test_users.Count) { var old_test_users = new HashSet <int>(test_users); var new_test_users = new int[num_test_users]; for (int i = 0; i < num_test_users; i++) { int random_index = MyMediaLite.Random.GetInstance().Next(old_test_users.Count - 1); new_test_users[i] = old_test_users.ElementAt(random_index); old_test_users.Remove(new_test_users[i]); } test_users = new_test_users; } // candidate items if (candidate_items_file != null) { candidate_items = item_mapping.ToInternalID(File.ReadLines(Path.Combine(data_dir, candidate_items_file)).ToArray()); } else if (all_items) { candidate_items = Enumerable.Range(0, item_mapping.InternalIDs.Max() + 1).ToArray(); } if (candidate_items != null) { eval_item_mode = CandidateItems.EXPLICIT; } else if (in_training_items) { eval_item_mode = CandidateItems.TRAINING; } else if (in_test_items) { eval_item_mode = CandidateItems.TEST; } else if (overlap_items) { eval_item_mode = CandidateItems.OVERLAP; } else { eval_item_mode = CandidateItems.UNION; } }); //Salvar arquivos List <string> linesToWrite = new List <string>(); for (int i = 0; i < training_data.UserMatrix.NumberOfRows; i++) { IList <int> columns = training_data.UserMatrix.GetEntriesByRow(i); for (int j = 0; j < columns.Count; j++) { StringBuilder line = new StringBuilder(); line.Append(i.ToString() + " " + columns[j].ToString()); linesToWrite.Add(line.ToString()); } } System.IO.File.WriteAllLines("training.data", linesToWrite.ToArray()); linesToWrite = new List <string>(); for (int i = 0; i < test_data.UserMatrix.NumberOfRows; i++) { IList <int> columns = test_data.UserMatrix.GetEntriesByRow(i); for (int j = 0; j < columns.Count; j++) { StringBuilder line = new StringBuilder(); line.Append(i.ToString() + " " + columns[j].ToString()); linesToWrite.Add(line.ToString()); } } System.IO.File.WriteAllLines("test.data", linesToWrite.ToArray()); /* * List<string> linesToWrite = new List<string>(); * for (int rowIndex = 0; rowIndex < training_data.AllItems.Count; rowIndex++) * { * * }*/ Console.Error.WriteLine(string.Format(CultureInfo.InvariantCulture, "loading_time {0,0:0.##}", loading_time.TotalSeconds)); Console.Error.WriteLine("memory {0}", Memory.Usage); }
/// <summary>Evaluate an iterative recommender on the folds of a dataset split, display results on STDOUT</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a positive-only feedback dataset split</param> /// <param name="test_users">a collection of integers with all test users</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <param name="max_iter">the maximum number of iterations</param> /// <param name="find_iter">the report interval</param> /// <param name="show_fold_results">if set to true to print per-fold results to STDERR</param> static public void DoRatingBasedRankingIterativeCrossValidation( this RatingPredictor recommender, ISplit <IRatings> split, IList <int> test_users, IList <int> candidate_items, CandidateItems candidate_item_mode, RepeatedEvents repeated_events, uint max_iter, uint find_iter = 1, bool show_fold_results = false) { if (!(recommender is IIterativeModel)) { throw new ArgumentException("recommender must be of type IIterativeModel"); } var split_recommenders = new RatingPredictor[split.NumberOfFolds]; var iterative_recommenders = new IIterativeModel[split.NumberOfFolds]; var fold_results = new ItemRecommendationEvaluationResults[split.NumberOfFolds]; // initial training and evaluation Parallel.For(0, (int)split.NumberOfFolds, i => { try { split_recommenders[i] = (RatingPredictor)recommender.Clone(); // to avoid changes in recommender split_recommenders[i].Ratings = split.Train[i]; split_recommenders[i].Train(); iterative_recommenders[i] = (IIterativeModel)split_recommenders[i]; var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[i]); var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[i]); fold_results[i] = Items.Evaluate(split_recommenders[i], test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode, repeated_events); if (show_fold_results) { Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, iterative_recommenders[i].NumIter); } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), iterative_recommenders[0].NumIter); // iterative training and evaluation for (int it = (int)iterative_recommenders[0].NumIter + 1; it <= max_iter; it++) { Parallel.For(0, (int)split.NumberOfFolds, i => { try { iterative_recommenders[i].Iterate(); if (it % find_iter == 0) { var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[i]); var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[i]); fold_results[i] = Items.Evaluate(split_recommenders[i], test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode, repeated_events); if (show_fold_results) { Console.WriteLine("fold {0} {1} iteration {2}", i, fold_results, it); } } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); Console.WriteLine("{0} iteration {1}", new ItemRecommendationEvaluationResults(fold_results), it); } }
/// <summary>Evaluation for rankings of items</summary> /// <remarks> /// User-item combinations that appear in both sets are ignored for the test set, and thus in the evaluation, /// except the boolean argument repeated_events is set. /// /// The evaluation measures are listed in the Measures property. /// Additionally, 'num_users' and 'num_items' report the number of users that were used to compute the results /// and the number of items that were taken into account. /// /// Literature: /// <list type="bullet"> /// <item><description> /// C. Manning, P. Raghavan, H. Schütze: Introduction to Information Retrieval, Cambridge University Press, 2008 /// </description></item> /// </list> /// /// On multi-core/multi-processor systems, the routine tries to use as many cores as possible, /// which should to an almost linear speed-up. /// </remarks> /// <param name="recommender">item recommender</param> /// <param name="test">test cases</param> /// <param name="training">training data</param> /// <param name="test_users">a list of integers with all test users; if null, use all users in the test cases</param> /// <param name="candidate_items">a list of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="repeated_events">allow repeated events in the evaluation (i.e. items accessed by a user before may be in the recommended list)</param> /// <returns>a dictionary containing the evaluation results (default is false)</returns> public static ItemRecommendationEvaluationResults Evaluate( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList<int> test_users = null, IList<int> candidate_items = null, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool repeated_events = false) { switch (candidate_item_mode) { case CandidateItems.TRAINING: candidate_items = training.AllItems; break; case CandidateItems.TEST: candidate_items = test.AllItems; break; case CandidateItems.OVERLAP: candidate_items = new List<int>(test.AllItems.Intersect(training.AllItems)); break; case CandidateItems.UNION: candidate_items = new List<int>(test.AllItems.Union(training.AllItems)); break; } if (candidate_items == null) throw new ArgumentNullException("candidate_items"); if (test_users == null) test_users = test.AllUsers; int num_users = 0; var result = new ItemRecommendationEvaluationResults(); // make sure that UserMatrix is completely initialized before entering parallel code var training_user_matrix = training.UserMatrix; var test_user_matrix = test.UserMatrix; Parallel.ForEach(test_users, user_id => { try { var correct_items = new HashSet<int>(test_user_matrix[user_id]); correct_items.IntersectWith(candidate_items); // the number of items that will be used for this user var candidate_items_in_train = training_user_matrix[user_id] == null ? new HashSet<int>() : new HashSet<int>(training_user_matrix[user_id]); candidate_items_in_train.IntersectWith(candidate_items); int num_eval_items = candidate_items.Count - (repeated_events ? 0 : candidate_items_in_train.Count()); // skip all users that have 0 or #candidate_items test items if (correct_items.Count == 0) return; if (num_eval_items == correct_items.Count) return; IList<int> prediction_list = recommender.PredictItems(user_id, candidate_items); if (prediction_list.Count != candidate_items.Count) throw new Exception("Not all items have been ranked."); ICollection<int> ignore_items = (repeated_events || training_user_matrix[user_id] == null) ? new int[0] : training_user_matrix[user_id]; double auc = AUC.Compute(prediction_list, correct_items, ignore_items); double map = PrecisionAndRecall.AP(prediction_list, correct_items, ignore_items); double ndcg = NDCG.Compute(prediction_list, correct_items, ignore_items); double rr = ReciprocalRank.Compute(prediction_list, correct_items, ignore_items); var positions = new int[] { 3, 5, 10 }; // DH: added for p@3 & r@3 var prec = PrecisionAndRecall.PrecisionAt(prediction_list, correct_items, ignore_items, positions); var recall = PrecisionAndRecall.RecallAt(prediction_list, correct_items, ignore_items, positions); // thread-safe incrementing lock (result) { num_users++; result["AUC"] += (float)auc; result["MAP"] += (float)map; result["NDCG"] += (float)ndcg; result["MRR"] += (float)rr; result["prec@3"] += (float)prec[3]; result["prec@5"] += (float)prec[5]; result["prec@10"] += (float)prec[10]; result["recall@3"] += (float)recall[3]; result["recall@5"] += (float)recall[5]; result["recall@10"] += (float)recall[10]; } if (num_users % 1000 == 0) Console.Error.Write("."); if (num_users % 60000 == 0) Console.Error.WriteLine(); } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); Console.Error.WriteLine("===> ERROR: user_id=" + user_id); Console.Error.WriteLine("===> ERROR: training_user_matrix[user_id]=" + training_user_matrix[user_id]); throw e; } }); foreach (string measure in Measures) result[measure] /= num_users; result["num_users"] = num_users; result["num_lists"] = num_users; result["num_items"] = candidate_items.Count; return result; }
// TODO consider micro- (by item) and macro-averaging (by user, the current thing); repeated events /// <summary>Online evaluation for rankings of items</summary> /// <remarks> /// </remarks> /// <param name="recommender">the item recommender to be evaluated</param> /// <param name="test">test cases</param> /// <param name="training">training data (must be connected to the recommender's training data)</param> /// <param name="test_users">a list of all test user IDs</param> /// <param name="candidate_items">a list of all candidate item IDs</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <returns>a dictionary containing the evaluation results (averaged by user)</returns> public static ItemRecommendationEvaluationResults EvaluateOnline( this IRecommender recommender, IPosOnlyFeedback test, IPosOnlyFeedback training, IList<int> test_users, IList<int> candidate_items, CandidateItems candidate_item_mode) { var incremental_recommender = recommender as IIncrementalItemRecommender; if (incremental_recommender == null) throw new ArgumentException("recommender must be of type IIncrementalItemRecommender"); // prepare candidate items once to avoid recreating them switch (candidate_item_mode) { case CandidateItems.TRAINING: candidate_items = training.AllItems; break; case CandidateItems.TEST: candidate_items = test.AllItems; break; case CandidateItems.OVERLAP: candidate_items = new List<int>(test.AllItems.Intersect(training.AllItems)); break; case CandidateItems.UNION: candidate_items = new List<int>(test.AllItems.Union(training.AllItems)); break; } candidate_item_mode = CandidateItems.EXPLICIT; // for better handling, move test data points into arrays var users = new int[test.Count]; var items = new int[test.Count]; int pos = 0; foreach (int user_id in test.UserMatrix.NonEmptyRowIDs) foreach (int item_id in test.UserMatrix[user_id]) { users[pos] = user_id; items[pos] = item_id; pos++; } // random order of the test data points // TODO chronological order var random_index = new int[test.Count]; for (int index = 0; index < random_index.Length; index++) random_index[index] = index; random_index.Shuffle(); var results_by_user = new Dictionary<int, ItemRecommendationEvaluationResults>(); int num_lists = 0; foreach (int index in random_index) { if (test_users.Contains(users[index]) && candidate_items.Contains(items[index])) { // evaluate user var current_test = new PosOnlyFeedback<SparseBooleanMatrix>(); current_test.Add(users[index], items[index]); var current_result = Items.Evaluate(recommender, current_test, training, current_test.AllUsers, candidate_items, candidate_item_mode); if (current_result["num_users"] == 1) if (results_by_user.ContainsKey(users[index])) { foreach (string measure in Items.Measures) results_by_user[users[index]][measure] += current_result[measure]; results_by_user[users[index]]["num_items"]++; num_lists++; } else { results_by_user[users[index]] = current_result; results_by_user[users[index]]["num_items"] = 1; results_by_user[users[index]].Remove("num_users"); } } // update recommender incremental_recommender.AddFeedback(users[index], items[index]); } var results = new ItemRecommendationEvaluationResults(); foreach (int u in results_by_user.Keys) foreach (string measure in Items.Measures) results[measure] += results_by_user[u][measure] / results_by_user[u]["num_items"]; foreach (string measure in Items.Measures) results[measure] /= results_by_user.Count; results["num_users"] = results_by_user.Count; results["num_items"] = candidate_items.Count; results["num_lists"] = num_lists; return results; }
static void LoadData() { TimeSpan loading_time = Wrap.MeasureTime(delegate() { // training data training_file = Path.Combine(data_dir, training_file); training_data = double.IsNaN(rating_threshold) ? ItemData.Read(training_file, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE) : ItemDataRatingThreshold.Read(training_file, rating_threshold, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE); // user attributes if (user_attributes_file != null) user_attributes = AttributeData.Read(Path.Combine(data_dir, user_attributes_file), user_mapping); if (recommender is IUserAttributeAwareRecommender) ((IUserAttributeAwareRecommender)recommender).UserAttributes = user_attributes; // item attributes if (item_attributes_file != null) item_attributes = AttributeData.Read(Path.Combine(data_dir, item_attributes_file), item_mapping); if (recommender is IItemAttributeAwareRecommender) ((IItemAttributeAwareRecommender)recommender).ItemAttributes = item_attributes; // user relation if (recommender is IUserRelationAwareRecommender) { ((IUserRelationAwareRecommender)recommender).UserRelation = RelationData.Read(Path.Combine(data_dir, user_relations_file), user_mapping); Console.WriteLine("relation over {0} users", ((IUserRelationAwareRecommender)recommender).NumUsers); } // item relation if (recommender is IItemRelationAwareRecommender) { ((IItemRelationAwareRecommender)recommender).ItemRelation = RelationData.Read(Path.Combine(data_dir, item_relations_file), item_mapping); Console.WriteLine("relation over {0} items", ((IItemRelationAwareRecommender)recommender).NumItems); } // user groups if (user_groups_file != null) { group_to_user = RelationData.Read(Path.Combine(data_dir, user_groups_file), user_mapping); // assumption: user and user group IDs are disjoint user_groups = group_to_user.NonEmptyRowIDs; Console.WriteLine("{0} user groups", user_groups.Count); } // test data if (test_ratio == 0) { if (test_file != null) { test_file = Path.Combine(data_dir, test_file); test_data = double.IsNaN(rating_threshold) ? ItemData.Read(test_file, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE) : ItemDataRatingThreshold.Read(test_file, rating_threshold, user_mapping, item_mapping, file_format == ItemDataFileFormat.IGNORE_FIRST_LINE); } } else { var split = new PosOnlyFeedbackSimpleSplit<PosOnlyFeedback<SparseBooleanMatrix>>(training_data, test_ratio); training_data = split.Train[0]; test_data = split.Test[0]; } if (group_method == "GroupsAsUsers") { Console.WriteLine("group recommendation strategy: {0}", group_method); // TODO verify what is going on here //var training_data_group = new PosOnlyFeedback<SparseBooleanMatrix>(); // transform groups to users foreach (int group_id in group_to_user.NonEmptyRowIDs) foreach (int user_id in group_to_user[group_id]) foreach (int item_id in training_data.UserMatrix.GetEntriesByRow(user_id)) training_data.Add(group_id, item_id); // add the users that do not belong to groups //training_data = training_data_group; // transform groups to users var test_data_group = new PosOnlyFeedback<SparseBooleanMatrix>(); foreach (int group_id in group_to_user.NonEmptyRowIDs) foreach (int user_id in group_to_user[group_id]) foreach (int item_id in test_data.UserMatrix.GetEntriesByRow(user_id)) test_data_group.Add(group_id, item_id); test_data = test_data_group; group_method = null; // deactivate s.t. the normal eval routines are used } if (user_prediction) { // swap file names for test users and candidate items var ruf = test_users_file; var rif = candidate_items_file; test_users_file = rif; candidate_items_file = ruf; // swap user and item mappings var um = user_mapping; var im = item_mapping; user_mapping = im; item_mapping = um; // transpose training and test data training_data = training_data.Transpose(); // transpose test data if (test_data != null) test_data = test_data.Transpose(); } if (recommender is MyMediaLite.ItemRecommendation.ItemRecommender) ((ItemRecommender)recommender).Feedback = training_data; // test users if (test_users_file != null) test_users = user_mapping.ToInternalID( File.ReadLines(Path.Combine(data_dir, test_users_file)).ToArray() ); else test_users = test_data != null ? test_data.AllUsers : training_data.AllUsers; // if necessary, perform user sampling if (num_test_users > 0 && num_test_users < test_users.Count) { var old_test_users = new HashSet<int>(test_users); var new_test_users = new int[num_test_users]; for (int i = 0; i < num_test_users; i++) { int random_index = MyMediaLite.Util.Random.GetInstance().Next(old_test_users.Count - 1); new_test_users[i] = old_test_users.ElementAt(random_index); old_test_users.Remove(new_test_users[i]); } test_users = new_test_users; } // candidate items if (candidate_items_file != null) candidate_items = item_mapping.ToInternalID( File.ReadLines(Path.Combine(data_dir, candidate_items_file)).ToArray() ); else if (all_items) candidate_items = Enumerable.Range(0, item_mapping.InternalIDs.Max() + 1).ToArray(); if (candidate_items != null) eval_item_mode = CandidateItems.EXPLICIT; else if (in_training_items) eval_item_mode = CandidateItems.TRAINING; else if (in_test_items) eval_item_mode = CandidateItems.TEST; else if (overlap_items) eval_item_mode = CandidateItems.OVERLAP; else eval_item_mode = CandidateItems.UNION; }); Console.Error.WriteLine(string.Format(CultureInfo.InvariantCulture, "loading_time {0,0:0.##}", loading_time.TotalSeconds)); Console.Error.WriteLine("memory {0}", Memory.Usage); }
/// <summary>Evaluate on the folds of a dataset split</summary> /// <param name="recommender">an item recommender</param> /// <param name="split">a dataset split</param> /// <param name="candidate_items">a collection of integers with all candidate items</param> /// <param name="candidate_item_mode">the mode used to determine the candidate items</param> /// <param name="compute_fit">if set to true measure fit on the training data as well</param> /// <param name="show_results">set to true to print results to STDERR</param> /// <returns>a dictionary containing the average results over the different folds of the split</returns> static public EvaluationResults DoRatingBasedRankingCrossValidation( this RatingPredictor recommender, ISplit <IRatings> split, IList <int> candidate_items, CandidateItems candidate_item_mode = CandidateItems.OVERLAP, bool compute_fit = false, bool show_results = false) { var avg_results = new ItemRecommendationEvaluationResults(); Parallel.For(0, (int)split.NumberOfFolds, fold => { try { var split_recommender = (RatingPredictor)recommender.Clone(); // avoid changes in recommender split_recommender.Ratings = split.Train[fold]; split_recommender.Train(); var test_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Test[fold]); var training_data_posonly = new PosOnlyFeedback <SparseBooleanMatrix>(split.Train[fold]); IList <int> test_users = test_data_posonly.AllUsers; var fold_results = Items.Evaluate(split_recommender, test_data_posonly, training_data_posonly, test_users, candidate_items, candidate_item_mode); if (compute_fit) { fold_results["fit"] = (float)split_recommender.ComputeFit(); } // thread-safe stats lock (avg_results) foreach (var key in fold_results.Keys) { if (avg_results.ContainsKey(key)) { avg_results[key] += fold_results[key]; } else { avg_results[key] = fold_results[key]; } } if (show_results) { Console.Error.WriteLine("fold {0} {1}", fold, fold_results); } } catch (Exception e) { Console.Error.WriteLine("===> ERROR: " + e.Message + e.StackTrace); throw; } }); foreach (var key in Items.Measures) { avg_results[key] /= split.NumberOfFolds; } avg_results["num_users"] /= split.NumberOfFolds; avg_results["num_items"] /= split.NumberOfFolds; return(avg_results); }